# llama.cpp 0.3 release fixes NaN bug and adds new binaries

> The open‑source LLaMA inference library updated on Nov 19, 2026 with a fix for sticky NaN values and new pre‑built binaries for macOS, iOS and Linux.

Oossa · 2026-11-19 · https://oossa.com/en/llama-cpp-0-3-release-fixes-nan-bug-and-adds-new-binaries

The ggml‑org team released llama.cpp version b12529 on Nov 19, 2026. The update fixes a sticky NaN issue that could corrupt the recurrent state in some models. It also adds pre‑compiled binaries for Apple Silicon, Intel macOS, iOS, and several Ubuntu builds, including support for newer hardware like s390x and Vulkan. The binaries are available for download from the GitHub release page.

## The facts

- Release version b12529 published on 2026-11-19
- Fixes sticky NaN in recurrent state; new binaries for macOS arm64, macOS x64, iOS, Ubuntu x64, arm64, s390x

## Why it matters

Developers can now run LLaMA models on more platforms without building from source, and the NaN fix improves stability for long‑running inference.

## Sources & references

1. [ggml-org/llama.cpp b12529](https://github.com/ggml-org/llama.cpp/releases/tag/b12529) – llama.cpp, 2026-11-19

Last updated: 2026-11-19
